clay-borg/games/ground/src/bot.rs

//! Bots — the kernel's first non-scenario consumer (CB-WP-0008 T01).
//!
//! Everything here drives `GroundState` through the ordinary
//! `Aggregate::validate` → `fold` path. Nothing in this module reaches
//! into state to mutate it: a bot that did would prove nothing about the
//! kernel, which is the whole reason INTENT wants a *second* consumer.
//!
//! **Stated limit on [`legal_commands`].** It is complete only up to the
//! candidate shapes it enumerates. Every candidate it yields is legal —
//! each one is filtered through `validate` — but a command shape this
//! function forgets to construct is invisible, and nothing here detects
//! that. Under-generation is the failure mode; it would show up as a bot
//! that never uses a rule, not as an error.

use crate::{
    Action, DarvoTarget, GroundChoice, GroundCommand, GroundMode, GroundState, Relation, RoundStep,
    Selection, SupportResponse,
};
use cb_kernel::{Actor, Aggregate, ChaChaRng, KernelRng, PlayerId, Rejection, Seed};

/// A seat's decision when offered the legal commands available to it.
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub enum Choice {
    /// Index into the offered slice.
    Command(usize),
    /// Decline to act. Legal only where the driver says it is.
    Pass,
}

/// How one seat decides. Implementors see the whole state — bots are not
/// the place to enforce hidden information; that is K13's projection, and
/// [`crate::bot`] deliberately does not pretend otherwise.
pub trait Policy {
    fn name(&self) -> &'static str;

    /// Pick one of `legal`, or [`Choice::Pass`] when `may_pass`.
    ///
    /// `legal` is never empty when this is called: the driver treats an
    /// empty legal set at an obligatory point as [`BotError::NoLegalMove`]
    /// rather than asking a policy to invent a move.
    fn choose(
        &mut self,
        state: &GroundState,
        seat: PlayerId,
        legal: &[GroundCommand],
        may_pass: bool,
    ) -> Choice;
}

/// Every way a bot game can fail. All of them are **loud**: the failure a
/// bot driver must never have is the silent one, where a stalled game is
/// indistinguishable from a finished one.
#[derive(Debug, Clone, PartialEq, Eq)]
pub enum BotError {
    /// A seat had to act and had nothing legal to do.
    NoLegalMove {
        seat: PlayerId,
        round: u8,
        step: RoundStep,
    },
    /// A policy passed where passing is not allowed.
    PassedWhenObligatory {
        seat: PlayerId,
        round: u8,
        step: RoundStep,
    },
    /// A policy returned an index outside the offered slice. Not clamped:
    /// clamping would turn a broken policy into a quietly playing one.
    IllegalChoice {
        seat: PlayerId,
        index: usize,
        offered: usize,
    },
    /// A command the driver generated was rejected. Since generation is
    /// validate-filtered, this means state moved underneath it.
    Rejected {
        seat: Option<PlayerId>,
        command: String,
        rejection: Rejection,
    },
    /// The game did not progress. A bot that loops forever looks like a
    /// bot that is working; this is the guard that makes it look broken.
    Stalled { round: u8, detail: String },
    /// Fewer policies than seats.
    NoPolicy { seat: PlayerId },
}

impl core::fmt::Display for BotError {
    fn fmt(&self, f: &mut core::fmt::Formatter<'_>) -> core::fmt::Result {
        match self {
            BotError::NoLegalMove { seat, round, step } => write!(
                f,
                "player {seat} has no legal command in round {round} {step:?}"
            ),
            BotError::PassedWhenObligatory { seat, round, step } => write!(
                f,
                "player {seat} passed in round {round} {step:?}, where an action is required"
            ),
            BotError::IllegalChoice {
                seat,
                index,
                offered,
            } => write!(
                f,
                "player {seat}'s policy chose index {index} of {offered} offered commands"
            ),
            BotError::Rejected {
                seat,
                command,
                rejection,
            } => write!(f, "{seat:?} issuing {command} was rejected: {rejection}"),
            BotError::Stalled { round, detail } => {
                write!(f, "game stalled in round {round}: {detail}")
            }
            BotError::NoPolicy { seat } => write!(f, "no policy for seat {seat}"),
        }
    }
}

/// What a completed bot game produced.
#[derive(Debug, Clone)]
pub struct BotGame {
    pub state: GroundState,
    /// Every event, in order — the log a replay bundle would carry.
    pub events: Vec<crate::GroundEvent>,
    /// Every accepted command, in order, as issued. `record::to_step`
    /// turns these into a scenario a replay bundle can carry.
    pub steps: Vec<(Actor, GroundCommand)>,
    /// Commands accepted, player and system alike.
    pub commands: usize,
    pub rounds: u8,
}

// ------------------------------------------------------------ generation

/// Every legal command `seat` may issue right now, in canonical order.
///
/// Candidates are constructed then filtered through `validate`, so the
/// game rules stay in one place. See the module note for what this
/// cannot tell you.
pub fn legal_commands(state: &GroundState, seat: PlayerId) -> Vec<GroundCommand> {
    let seats: Vec<PlayerId> = state.players.keys().copied().collect();
    let problems: Vec<u32> = state.problems.keys().copied().collect();
    let mut candidates: Vec<GroundCommand> = Vec::new();

    match state.step {
        RoundStep::Select => {
            for action in [
                Action::Investigate,
                Action::Solve,
                Action::Support,
                Action::Attack,
                Action::Ground,
            ] {
                match action {
                    Action::Investigate | Action::Solve => {
                        for problem in &problems {
                            candidates.push(GroundCommand::SelectAction {
                                action,
                                target: None,
                                problem: Some(*problem),
                            });
                        }
                    }
                    Action::Support | Action::Attack => {
                        for other in &seats {
                            candidates.push(GroundCommand::SelectAction {
                                action,
                                target: Some(*other),
                                problem: None,
                            });
                        }
                    }
                    Action::Ground => candidates.push(GroundCommand::SelectAction {
                        action,
                        target: None,
                        problem: None,
                    }),
                }
            }
            candidates.push(GroundCommand::SpendFreedom);
        }
        RoundStep::Reveal => {
            candidates.push(GroundCommand::ChooseGroundMode {
                mode: GroundMode::Gr,
                choice: None,
            });
            for problem in &problems {
                for choice in [
                    GroundChoice::RestoreProblem { problem: *problem },
                    GroundChoice::ProtectProblem { problem: *problem },
                ] {
                    candidates.push(GroundCommand::ChooseGroundMode {
                        mode: GroundMode::Ou,
                        choice: Some(choice),
                    });
                }
            }
            for other in &seats {
                candidates.push(GroundCommand::ChooseGroundMode {
                    mode: GroundMode::Ou,
                    choice: Some(GroundChoice::CancelAttack { attacker: *other }),
                });
                for choice in [
                    GroundChoice::RemoveBlame { owner: *other },
                    GroundChoice::BreakRelation { with: *other },
                ] {
                    candidates.push(GroundCommand::ChooseGroundMode {
                        mode: GroundMode::Nd,
                        choice: Some(choice),
                    });
                }
            }
            candidates.push(GroundCommand::ChooseGroundMode {
                mode: GroundMode::Nd,
                choice: Some(GroundChoice::RejectReverse),
            });

            for response in [
                SupportResponse::AcceptBond,
                SupportResponse::DeclineBond,
                SupportResponse::FlipToBond,
                SupportResponse::BreakRivalry,
            ] {
                candidates.push(GroundCommand::RespondToSupport { response });
            }

            for problem in &problems {
                candidates.push(GroundCommand::ChooseDarvoTarget {
                    target: DarvoTarget {
                        problem: Some(*problem),
                        player: None,
                    },
                });
            }
            for other in &seats {
                candidates.push(GroundCommand::ChooseDarvoTarget {
                    target: DarvoTarget {
                        problem: None,
                        player: Some(*other),
                    },
                });
            }
        }
        // GR-R06/R08: system-driven; a seat has nothing to issue.
        RoundStep::Resolve | RoundStep::End => {}
    }

    candidates
        .into_iter()
        .filter(|cmd| state.validate(Actor::Player(seat), cmd).is_ok())
        .collect()
}

// --------------------------------------------------------------- policies

/// Uniform over the legal commands, from a seeded kernel RNG so a bot
/// game stays deterministic and replayable (K8).
pub struct RandomPolicy {
    rng: ChaChaRng,
    /// Probability, in sixteenths, of taking an optional action rather
    /// than passing. Fixed rather than tunable: a knob here would be a
    /// parameter nobody has a second use for.
    act_in_16: u32,
}

impl RandomPolicy {
    pub fn new(seed: u64) -> Self {
        Self {
            rng: ChaChaRng::from_seed(Seed(seed)),
            act_in_16: 12,
        }
    }
}

impl Policy for RandomPolicy {
    fn name(&self) -> &'static str {
        "random"
    }

    fn choose(
        &mut self,
        _state: &GroundState,
        _seat: PlayerId,
        legal: &[GroundCommand],
        may_pass: bool,
    ) -> Choice {
        if may_pass && self.rng.draw(16) >= self.act_in_16 {
            return Choice::Pass;
        }
        Choice::Command(self.rng.draw(legal.len() as u32) as usize)
    }
}

/// A stated heuristic, in priority order:
///
/// 1. **Get out from under the stress gate.** At Stress ≥ 4 only ATTACK
///    and GROUND are selectable (GR-R03), so GROUND—GR (−2 Stress) is
///    worth more than anything else on the board.
/// 2. **SOLVE** a face-up Problem — the only action that adds to the
///    shared total the outcome is scored against (GR-E01).
/// 3. **INVESTIGATE** — turns a hidden Problem face up, which is what
///    makes a later SOLVE possible.
/// 4. **SUPPORT** — Bonds are the cheapest relationship and feed GR-E04.
/// 5. **GROUND**, then **ATTACK** last: attacking raises another seat's
///    Stress, which under SHARED GROUND scoring costs the group.
///
/// Deterministic by construction: ties break toward the earlier candidate
/// in canonical order, so two runs of a greedy game are identical without
/// needing a seed.
#[derive(Debug, Default, Clone, Copy)]
pub struct GreedyPolicy;

impl GreedyPolicy {
    fn rank(state: &GroundState, seat: PlayerId, cmd: &GroundCommand) -> i32 {
        let gated = state
            .players
            .get(&seat)
            .is_some_and(|p| p.stress >= 4 && !p.freedom_gate_lifted);
        match cmd {
            GroundCommand::SelectAction {
                action, problem, ..
            } => match action {
                Action::Ground if gated => 100,
                // GR-A13 admits SOLVE against an already-claimed Problem
                // (it names only "face-up, non-Denied"), and resolution
                // then does nothing — the action is silently wasted. The
                // policy avoids it; the *rule* is left alone, and the gap
                // is recorded rather than patched here.
                Action::Solve
                    if problem
                        .and_then(|p| state.problems.get(&p))
                        .is_some_and(|p| p.claimed_by.is_some()) =>
                {
                    5
                }
                Action::Solve => 90,
                Action::Investigate => 80,
                Action::Support => 70,
                Action::Ground => 60,
                Action::Attack => 10,
            },
            // Only worth spending when the gate is actually biting.
            GroundCommand::SpendFreedom => {
                if gated {
                    95
                } else {
                    0
                }
            }
            GroundCommand::ChooseGroundMode { mode, choice } => match (mode, choice) {
                // GR-A10: the only mode that lowers our own Stress.
                (GroundMode::Gr, _) => 90,
                (_, Some(GroundChoice::RestoreProblem { .. })) => 80,
                (_, Some(GroundChoice::CancelAttack { .. })) => 75,
                (_, Some(GroundChoice::ProtectProblem { .. })) => 70,
                (_, Some(GroundChoice::RemoveBlame { .. })) => 65,
                _ => 40,
            },
            // GR-L02/A05: take the Bond wherever one is on offer.
            GroundCommand::RespondToSupport { response } => match response {
                SupportResponse::AcceptBond | SupportResponse::FlipToBond => 90,
                SupportResponse::BreakRivalry => 50,
                SupportResponse::DeclineBond => 10,
            },
            GroundCommand::ChooseDarvoTarget { .. } => 50,
            GroundCommand::Reveal | GroundCommand::Resolve | GroundCommand::EndRound => -1,
        }
    }
}

impl Policy for GreedyPolicy {
    fn name(&self) -> &'static str {
        "greedy"
    }

    fn choose(
        &mut self,
        state: &GroundState,
        seat: PlayerId,
        legal: &[GroundCommand],
        _may_pass: bool,
    ) -> Choice {
        let mut best = 0usize;
        let mut best_rank = i32::MIN;
        for (i, cmd) in legal.iter().enumerate() {
            let rank = Self::rank(state, seat, cmd);
            if rank > best_rank {
                best_rank = rank;
                best = i;
            }
        }
        Choice::Command(best)
    }
}

// ----------------------------------------------------------------- driver

/// Rounds the driver will run before declaring a stall. GR-R09 ends the
/// game after five; anything past that is a defect, not a long game.
const MAX_ROUNDS: u8 = 20;

/// Attempts a single seat gets within one step. Bounded so a policy that
/// keeps choosing non-advancing commands fails instead of spinning.
const MAX_ATTEMPTS: usize = 8;

/// Drive `state` to `GameEnded` with one policy per seat.
///
/// Seats are matched to `policies` by index: `PlayerId(n)` gets
/// `policies[n]`.
pub fn play<'a>(
    state: GroundState,
    policies: &mut [Box<dyn Policy + 'a>],
) -> Result<BotGame, BotError> {
    play_journaled(state, policies, None)
}

/// The same, appending every applied command and its events to `journal`
/// as it goes, for a caller rendering the game while it runs.
pub fn play_journaled<'a>(
    mut state: GroundState,
    policies: &mut [Box<dyn Policy + 'a>],
    journal: Option<Journal>,
) -> Result<BotGame, BotError> {
    let mut log = Log {
        journal,
        ..Log::default()
    };

    let seats: Vec<PlayerId> = state.players.keys().copied().collect();
    for seat in &seats {
        if policies.get(seat.0 as usize).is_none() {
            return Err(BotError::NoPolicy { seat: *seat });
        }
    }

    let mut rounds = 0u8;
    while state.outcome.is_none() {
        rounds += 1;
        if rounds > MAX_ROUNDS {
            return Err(BotError::Stalled {
                round: rounds,
                detail: format!("no outcome after {MAX_ROUNDS} rounds"),
            });
        }
        let round = state.round;

        // GR-R02 — Select. Every seat must end this step with a selection.
        for seat in &seats {
            let mut attempts = 0;
            while !state.selections.contains_key(seat) {
                attempts += 1;
                if attempts > MAX_ATTEMPTS {
                    return Err(BotError::Stalled {
                        round,
                        detail: format!("player {seat} never selected an action"),
                    });
                }
                step_seat(&mut state, &mut log, policies, *seat, round, false)?;
            }
        }
        // Positive control: the driver must have done the work it claims.
        if state.selections.len() != state.players.len() {
            return Err(BotError::Stalled {
                round,
                detail: format!(
                    "Select ended with {} of {} selections",
                    state.selections.len(),
                    state.players.len()
                ),
            });
        }

        apply(&mut state, &mut log, Actor::System, &GroundCommand::Reveal)?;

        // GR-R05 — after Reveal: modes are obligatory for a seat that
        // revealed GROUND; Support responses and DARVO targets are not.
        for seat in &seats {
            let mut attempts = 0;
            loop {
                attempts += 1;
                if attempts > MAX_ATTEMPTS {
                    return Err(BotError::Stalled {
                        round,
                        detail: format!("player {seat} kept acting after Reveal"),
                    });
                }
                let obligatory = state
                    .selections
                    .get(seat)
                    .is_some_and(|s| s.action == Action::Ground)
                    && !state.ground_modes.contains_key(seat);
                if !obligatory && legal_commands(&state, *seat).is_empty() {
                    break;
                }
                if !step_seat(&mut state, &mut log, policies, *seat, round, !obligatory)? {
                    break;
                }
            }
        }

        apply(&mut state, &mut log, Actor::System, &GroundCommand::Resolve)?;
        apply(
            &mut state,
            &mut log,
            Actor::System,
            &GroundCommand::EndRound,
        )?;
    }

    Ok(BotGame {
        state,
        commands: log.steps.len(),
        events: log.events,
        steps: log.steps,
        rounds,
    })
}

/// One command and everything it produced.
///
/// **The empty case is the load-bearing one** (CB-WP-0018 T02). GR-A02's
/// resolver silently `continue`s when a SOLVE cannot be fulfilled, so a
/// player can select it three rounds running and change nothing. A journal
/// built only from events would show nothing for those and reproduce the
/// silence; keeping the command with an empty `events` is what lets a
/// reader say *"this happened and did nothing"*.
#[derive(Debug, Clone)]
pub struct Applied {
    pub actor: Actor,
    pub command: GroundCommand,
    pub events: Vec<crate::GroundEvent>,
}

/// A live account of a game in progress, shared with whoever is watching.
///
/// `BotGame.events` is the same information but only after `play` returns,
/// which is no use to a page rendered mid-game.
pub type Journal = std::rc::Rc<std::cell::RefCell<Vec<Applied>>>;

/// What the driver accumulates while a game runs.
#[derive(Default)]
struct Log {
    events: Vec<crate::GroundEvent>,
    steps: Vec<(Actor, GroundCommand)>,
    journal: Option<Journal>,
}

/// Offer one seat its legal commands and apply what the policy picks.
/// `Ok(false)` means the seat passed.
fn step_seat(
    state: &mut GroundState,
    log: &mut Log,
    policies: &mut [Box<dyn Policy + '_>],
    seat: PlayerId,
    round: u8,
    may_pass: bool,
) -> Result<bool, BotError> {
    let legal = legal_commands(state, seat);
    if legal.is_empty() {
        if may_pass {
            return Ok(false);
        }
        return Err(BotError::NoLegalMove {
            seat,
            round,
            step: state.step,
        });
    }
    let policy = policies
        .get_mut(seat.0 as usize)
        .ok_or(BotError::NoPolicy { seat })?;
    match policy.choose(state, seat, &legal, may_pass) {
        Choice::Pass if may_pass => Ok(false),
        Choice::Pass => Err(BotError::PassedWhenObligatory {
            seat,
            round,
            step: state.step,
        }),
        Choice::Command(i) => {
            let cmd = legal.get(i).ok_or(BotError::IllegalChoice {
                seat,
                index: i,
                offered: legal.len(),
            })?;
            apply(state, log, Actor::Player(seat), cmd)?;
            Ok(true)
        }
    }
}

fn apply(
    state: &mut GroundState,
    log: &mut Log,
    actor: Actor,
    cmd: &GroundCommand,
) -> Result<(), BotError> {
    let produced = state
        .validate(actor, cmd)
        .map_err(|rejection| BotError::Rejected {
            seat: match actor {
                Actor::Player(p) => Some(p),
                Actor::System => None,
            },
            command: format!("{cmd:?}"),
            rejection,
        })?;
    for event in &produced {
        state.fold(event);
    }
    if let Some(j) = &log.journal {
        j.borrow_mut().push(Applied {
            actor,
            command: cmd.clone(),
            events: produced.clone(),
        });
    }
    log.events.extend(produced);
    log.steps.push((actor, cmd.clone()));
    Ok(())
}

/// The selection a seat made this round, for callers that want to report
/// a game rather than only run one.
pub fn selection_of(state: &GroundState, seat: PlayerId) -> Option<Selection> {
    state.selections.get(&seat).copied()
}

/// Relation between two seats, for the same reason.
pub fn relation_of(state: &GroundState, a: PlayerId, b: PlayerId) -> Option<Relation> {
    state.relations.get(&crate::Pair::new(a, b)).copied()
}

#[cfg(all(test, feature = "scenarios"))]
mod tests {
    use super::*;
    use cb_events::state_hash_hex;
    use cb_game_runtime::{ScenarioGame, Setup};

    fn setup(players: u8, seed: u64) -> GroundState {
        GroundState::setup(
            &Setup {
                players,
                preset: format!("standard-{players}p"),
                patch: Default::default(),
            },
            seed,
        )
        .expect("preset")
    }

    fn policies(kind: &str, n: u8, seed: u64) -> Vec<Box<dyn Policy>> {
        (0..n)
            .map(|i| -> Box<dyn Policy> {
                match kind {
                    "greedy" => Box::new(GreedyPolicy),
                    _ => Box::new(RandomPolicy::new(seed + u64::from(i))),
                }
            })
            .collect()
    }

    /// Acceptance, first half: a 3-player all-bot game reaches GR-R09's
    /// end with every command legal (nothing was rejected — `play`
    /// returns `Err` on the first rejection).
    #[test]
    fn three_player_all_bot_game_reaches_game_ended() {
        for kind in ["random", "greedy"] {
            let game = play(setup(3, 42), &mut policies(kind, 3, 42))
                .unwrap_or_else(|e| panic!("{kind} bots: {e}"));
            let outcome = game.state.outcome.as_ref().expect("GR-R09 outcome");
            assert_eq!(game.state.round, 5, "{kind}: GR-R09 runs five rounds");
            assert!(
                game.events
                    .iter()
                    .any(|e| matches!(e, crate::GroundEvent::GameEnded { .. })),
                "{kind}: GameEnded must be in the log"
            );
            // Positive control: a driver that did nothing would still see
            // an outcome if the state arrived scored, so assert the work.
            assert!(
                game.commands > 3 * 5,
                "{kind}: only {} commands for a five-round game",
                game.commands
            );
            assert!(outcome.threshold > 0);
        }
    }

    /// Acceptance, second half (K8): same seed, identical state hash.
    #[test]
    fn same_seed_bot_games_are_hash_identical() {
        let a = play(setup(3, 7), &mut policies("random", 3, 7)).expect("run a");
        let b = play(setup(3, 7), &mut policies("random", 3, 7)).expect("run b");
        assert_eq!(
            state_hash_hex(&a.state),
            state_hash_hex(&b.state),
            "same-seed bot games must be hash-identical"
        );
        assert_eq!(a.events.len(), b.events.len());

        // And the seed must matter, or the equality above is vacuous —
        // two runs of a bot that ignores its RNG are also identical.
        let c = play(setup(3, 7), &mut policies("random", 3, 999)).expect("run c");
        assert_ne!(
            state_hash_hex(&a.state),
            state_hash_hex(&c.state),
            "different policy seeds must produce a different game"
        );
    }

    /// The positive control the workplan asks for: no legal move must be
    /// loud. A seat the game does not contain has no legal command, and
    /// the driver must say so rather than skip it.
    #[test]
    fn a_seat_with_no_legal_move_fails_loudly() {
        let mut state = setup(3, 1);
        let ghost = PlayerId(9);
        assert!(legal_commands(&state, ghost).is_empty());
        let mut ps = policies("greedy", 10, 0);
        let err = step_seat(&mut state, &mut Log::default(), &mut ps, ghost, 1, false)
            .expect_err("a seat with nothing legal must fail");
        assert!(matches!(err, BotError::NoLegalMove { seat, .. } if seat == ghost));
    }

    /// A policy that passes where it may not is a defect, not a turn.
    #[test]
    fn passing_when_obligatory_fails_loudly() {
        struct Passer;
        impl Policy for Passer {
            fn name(&self) -> &'static str {
                "passer"
            }
            fn choose(
                &mut self,
                _: &GroundState,
                _: PlayerId,
                _: &[GroundCommand],
                _: bool,
            ) -> Choice {
                Choice::Pass
            }
        }
        let mut ps: Vec<Box<dyn Policy>> = (0..3)
            .map(|_| Box::new(Passer) as Box<dyn Policy>)
            .collect();
        let err = play(setup(3, 3), &mut ps).expect_err("a passing bot must not finish a game");
        assert!(
            matches!(err, BotError::PassedWhenObligatory { .. }),
            "got {err}"
        );
    }

    /// An out-of-range index is not clamped to a legal move.
    #[test]
    fn an_out_of_range_choice_fails_loudly() {
        struct Wild;
        impl Policy for Wild {
            fn name(&self) -> &'static str {
                "wild"
            }
            fn choose(
                &mut self,
                _: &GroundState,
                _: PlayerId,
                legal: &[GroundCommand],
                _: bool,
            ) -> Choice {
                Choice::Command(legal.len())
            }
        }
        let mut ps: Vec<Box<dyn Policy>> =
            (0..3).map(|_| Box::new(Wild) as Box<dyn Policy>).collect();
        let err = play(setup(3, 3), &mut ps).expect_err("an out-of-range index must fail");
        assert!(matches!(err, BotError::IllegalChoice { .. }), "got {err}");
    }

    /// Every command a bot issues goes through `validate`. This asserts
    /// the generator agrees with the kernel rather than trusting it.
    #[test]
    fn every_generated_command_validates() {
        let state = setup(3, 11);
        let mut total = 0;
        for seat in state.players.keys().copied() {
            let legal = legal_commands(&state, seat);
            assert!(!legal.is_empty(), "seat {seat} has no opening move");
            for cmd in &legal {
                assert!(
                    state.validate(Actor::Player(seat), cmd).is_ok(),
                    "generated {cmd:?} is not legal for {seat}"
                );
            }
            total += legal.len();
        }
        assert!(
            total > 10,
            "only {total} legal opening commands across 3 seats"
        );
    }

    /// The greedy heuristic must actually be a heuristic — a policy that
    /// ranked everything equally would pick index 0 and still finish.
    #[test]
    fn greedy_prefers_solve_over_attack() {
        let state = setup(3, 5);
        let seat = PlayerId(0);
        let legal = legal_commands(&state, seat);
        let mut p = GreedyPolicy;
        let Choice::Command(i) = p.choose(&state, seat, &legal, false) else {
            panic!("greedy never passes");
        };
        match &legal[i] {
            GroundCommand::SelectAction { action, .. } => assert!(
                matches!(action, Action::Solve | Action::Investigate),
                "greedy opened with {action:?}"
            ),
            other => panic!("greedy opened with {other:?}"),
        }
    }
    /// HDN control: a bot game that "finishes" without touching the board
    /// is the failure this project keeps finding. Assert the game moved.
    #[test]
    fn a_bot_game_does_substantive_work() {
        for kind in ["random", "greedy"] {
            let g = play(setup(3, 42), &mut policies(kind, 3, 42)).expect("game");
            let mut counts: std::collections::BTreeMap<String, usize> =
                std::collections::BTreeMap::new();
            for e in &g.events {
                let text = format!("{e:?}");
                let kind = text.split([' ', '{']).next().unwrap_or("?").to_string();
                *counts.entry(kind).or_default() += 1;
            }
            println!(
                "{kind}: cmds={} events={} {counts:?}",
                g.commands,
                g.events.len()
            );
            println!("   outcome {:?}", g.state.outcome);

            // Every seat selected in every round (GR-R02 × GR-R09).
            assert_eq!(
                counts.get("ActionSelected").copied().unwrap_or(0),
                3 * 5,
                "{kind}: a five-round three-player game has fifteen selections"
            );
            // And resolution actually did something with them. A driver
            // that selected, revealed and ended without resolving would
            // pass every assertion above this line.
            let board_moved: usize = [
                "ProblemClaimed",
                "ProblemRevealed",
                "RelationFormed",
                "StressSet",
                "SolutionDrawn",
            ]
            .iter()
            .filter_map(|k| counts.get(*k))
            .sum();
            assert!(
                board_moved >= 4,
                "{kind}: only {board_moved} board-changing events in a whole game"
            );
        }
    }
    /// GR-O01 says 2–6 players. Every scenario in the corpus was
    /// 3-player until CB-WP-0008 T03, so the range was stated and tested
    /// at one point. This plays all five counts with both policies.
    #[test]
    fn every_seat_count_in_gr_o01_plays_to_the_end() {
        for players in 2..=6u8 {
            for kind in ["random", "greedy"] {
                let game = play(setup(players, 42), &mut policies(kind, players, 42))
                    .unwrap_or_else(|e| panic!("{players}p {kind}: {e}"));
                let outcome = game
                    .state
                    .outcome
                    .as_ref()
                    .unwrap_or_else(|| panic!("{players}p {kind}: no outcome"));
                assert_eq!(game.state.round, 5, "{players}p {kind}: GR-R09");
                assert_eq!(
                    outcome.personal.len(),
                    usize::from(players),
                    "{players}p {kind}: every seat must be scored"
                );
                // K8 at each boundary, not only at three seats.
                let again =
                    play(setup(players, 42), &mut policies(kind, players, 42)).expect("second run");
                assert_eq!(
                    cb_events::state_hash_hex(&game.state),
                    cb_events::state_hash_hex(&again.state),
                    "{players}p {kind}: same seed must reproduce"
                );
            }
        }
    }

    /// **A recorded finding, not a desired property.** With the standard
    /// preset's placeholder Problem values (value = priority), the total
    /// a game can possibly reach is below GR-E01's threshold at 2, 3 and
    /// 4 players — group success is unreachable regardless of play. Only
    /// 5–6p can clear its 9.
    ///
    /// This test pins the arithmetic so the gap cannot close silently.
    /// **It is expected to fail** when Problem values become scenario
    /// data (GR-S01 calls the current fixture a stand-in); the failure is
    /// the signal to delete it, not to re-tune it.
    #[test]
    fn the_standard_preset_cannot_reach_the_threshold_below_five_seats() {
        let mut report = Vec::new();
        for players in 2..=6u8 {
            let initial = setup(players, 42);
            let best: u32 = initial.problems.values().map(|p| u32::from(p.value)).sum();
            let threshold = play(initial, &mut policies("greedy", players, 42))
                .expect("game")
                .state
                .outcome
                .expect("outcome")
                .threshold;
            report.push(format!("{players}p best {best} vs threshold {threshold}"));
            if players < 5 {
                assert!(
                    best < threshold,
                    "{players}p: best {best} now reaches threshold {threshold} — \
                     the fixture changed, delete this test"
                );
            } else {
                assert!(
                    best >= threshold,
                    "{players}p: best {best} cannot reach threshold {threshold}"
                );
            }
        }
        println!("GR-E01 reachability: {}", report.join(", "));
    }
}